[OMPT]: Fix the order of implicit_task_end_events

[tcramer]

Summary

For implcit barriers in simple parallel for loops the order of the OMPT events was wrong. The barrier_{begin,end} events came after the implcit_task_end event for the implcit barrier at the end of the parallel region. This wrong, because the implicit task executes the barrier before ending. This patch fixes the order of the event: It will be triggerd now just before __kmp_pop_current_task_from_thread() is called.

[jmellorcrummey]

LGTM.

[jlpeyton]

LGTM as well

runtime/src/kmp_runtime.c

#if OMP_40_ENABLED
     if ( !master_th->th.th_teams_microtask || team->t.t_level > master_th->th.th_teams_level )
 #endif /* OMP_40_ENABLED */
     {
         /* Decrement our nested depth level */
         KMP_TEST_THEN_DEC32( (kmp_int32*) &root->r.r_in_parallel );
     }
     KMP_DEBUG_ASSERT( root->r.r_in_parallel >= 0 );
 
+#if OMPT_SUPPORT && OMPT_TRACE
+    if(ompt_enabled){
+        ompt_task_info_t *task_info = __ompt_get_taskinfo(0);
+        if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) {
+             ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)(
+               parallel_id, task_info->task_id);
+        }
+        task_info->frame.exit_runtime_frame = 0;
+        task_info->task_id = 0;
+    }
+#endif
+
     KF_TRACE( 10, ("__kmp_join_call1: T#%d, this_thread=%p team=%p\n",
                    0, master_th, team ) );
     __kmp_pop_current_task_from_thread( master_th );
 
 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
     //
     // Restore master thread's partition.
     //
#if OMPT_SUPPORT
             this_thr->th.ompt_thread_info.state = ompt_state_overhead;
         }
 #endif
 
         pteam = (kmp_team_t *(*))(& this_thr->th.th_team);
 
         /* have we been allocated? */
         if ( TCR_SYNC_PTR(*pteam) && !TCR_4(__kmp_global.g.g_done) ) {
+#if OMPT_SUPPORT
+            ompt_task_info_t *task_info;
+            if (ompt_enabled) {
+                task_info = __ompt_get_taskinfo(0);
+            }
+#endif
             /* we were just woken up, so run our new task */
             if ( TCR_SYNC_PTR((*pteam)->t.t_pkfn) != NULL ) {
                 int rc;
                 KA_TRACE(20, ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n",
                               gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid), (*pteam)->t.t_pkfn));
 
                 updateHWFPControl (*pteam);
 
 #if OMPT_SUPPORT
                 if (ompt_enabled) {
                     this_thr->th.ompt_thread_info.state = ompt_state_work_parallel;
                     // Initialize OMPT task id for implicit task.
                     int tid = __kmp_tid_from_gtid(gtid);
-                    (*pteam)->t.t_implicit_task_taskdata[tid].ompt_task_info.task_id =
-                    __ompt_task_id_new(tid);
+                    task_info->task_id = __ompt_task_id_new(tid);
                 }
 #endif
 
                 KMP_STOP_DEVELOPER_EXPLICIT_TIMER(USER_launch_thread_loop);
                 {
                     KMP_TIME_DEVELOPER_BLOCK(USER_worker_invoke);
                     rc = (*pteam)->t.t_invoke( gtid );
                 }
                 KMP_START_DEVELOPER_EXPLICIT_TIMER(USER_launch_thread_loop);
                 KMP_ASSERT( rc );
 
 #if OMPT_SUPPORT
                 if (ompt_enabled) {
                     /* no frame set while outside task */
-                    int tid = __kmp_tid_from_gtid(gtid);
-                    (*pteam)->t.t_implicit_task_taskdata[tid].ompt_task_info.frame.exit_runtime_frame = 0;
+                    task_info->frame.exit_runtime_frame = 0;
 
                     this_thr->th.ompt_thread_info.state = ompt_state_overhead;
                 }
 #endif
                 KMP_MB();
                 KA_TRACE(20, ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n",
                               gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid), (*pteam)->t.t_pkfn));
             }
             /* join barrier after parallel region */
             __kmp_join_barrier( gtid );
+#if OMPT_SUPPORT && OMPT_TRACE
+            if (ompt_enabled) {
+                if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) {
+                    int my_parallel_id = (*pteam)->t.ompt_team_info.parallel_id;
+                    ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)(
+                        my_parallel_id, task_info->task_id);
+                }
+                task_info->frame.exit_runtime_frame = 0;
+                task_info->task_id = 0;
+            }
+#endif 
         }
     }
     TCR_SYNC_PTR((intptr_t)__kmp_global.g.g_done);
 
 #if OMPT_SUPPORT
     if (ompt_enabled &&
         ompt_callbacks.ompt_callback(ompt_event_thread_end)) {
         __ompt_thread_end(ompt_thread_worker, gtid);
#endif
         rc = __kmp_invoke_microtask( (microtask_t) TCR_SYNC_PTR(team->t.t_pkfn),
                                      gtid, tid, (int) team->t.t_argc, (void **) team->t.t_argv
 #if OMPT_SUPPORT
                                      , exit_runtime_p
 #endif
                                      );
     }
 
-#if OMPT_SUPPORT && OMPT_TRACE
-    if (ompt_enabled) {
-        if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) {
-            ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)(
-                my_parallel_id, my_task_id);
-        }
-        // the implicit task is not dead yet, so we can't clear its task id here
-        team->t.t_implicit_task_taskdata[tid].ompt_task_info.frame.exit_runtime_frame = 0;
-    }
-#endif
-
 #if USE_ITT_BUILD
     if ( __itt_stack_caller_create_ptr ) {
         __kmp_itt_stack_callee_leave( (__itt_caller)team->t.t_stack_id ); // inform ittnotify about leaving user's code
     }
 #endif /* USE_ITT_BUILD */
     __kmp_run_after_invoked_task( gtid, tid, this_thr, team );
 
     return rc;